MSCCL++ v0.6.0

Highlight

• Improved NCCL API integration in MSCCL++ for better performance and usability
• Enhanced execution plan-based executor in MSCCL++
• Fixed several bugs to improve stability and reliability

What's Changed

• Add support for different vector sizes in multimem instructions by @roshandathathri in #332
• NCCL API Executor Integration by @caiomcbr in #331
• Fix missing import in executor test by @yzygitzh in #334
• bfloat16 support by @chhwang in #336
• Dynamically load libibverbs by @caiomcbr in #337
• Auto-tune vector sizes for NVLS allreduce6 by @roshandathathri in #338
• Make ibverbs optional at compile time by @chhwang in #340
• ProxyChannel Support in Executor by @caiomcbr in #342
• Support executors to send packets over ProxyChannel by @caiomcbr in #344
• Fix for ROCm 6.0 by @chhwang in #347
• Fix bug for construct sempaphore by @Binyang2014 in #341
• Add proxy channel related operations by @Binyang2014 in #351
• Add CI for rocm by @Binyang2014 in #346
• Tune threads per block for mscclpp executor by @Binyang2014 in #345
• Fix NPKit exit event offset by @yzygitzh in #356
• Use IB transport flags only when an IB device exists by @chhwang in #355
• Update ROCm CI by @chhwang in #357
• Fixing RegisterMemory Allocation for ProxyChannels by @caiomcbr in #353
• Fix NCCL API bugs by @chhwang in #363
• Perf optimization & support clipping by @chhwang in #364
• Fix copyright messages by @chhwang in #367
• [Doc] mscclpp docs by @Binyang2014 in #348
• Executor AllGather In-Place Support by @caiomcbr in #365
• Fix algo repo name by @Binyang2014 in #369
• Update docker image for cuda12.4 by @Binyang2014 in #370
• Fix in-place all-gather input buffer in executor_test by @yzygitzh in #372
• [docs] fix quickstart link by @jeffra in #374
• Add kernel-based verification for executor_test by @yzygitzh in #378
• Lazily create the context stream by @chhwang in #381
• Fixing Bug Const Offset in Execution Plan by @caiomcbr in #380
• Fix light load bug by @Binyang2014 in #379
• Small Adjust in Test Data AllGather at Executor Test by @caiomcbr in #384
• Fix missing packet parameter for executor by @yzygitzh in #385
• NVLS support for msccl++ executor by @Binyang2014 in #375
• Fix typo by @Binyang2014 in #389
• Improve CMake options by @chhwang in #376
• Fixing Message Boundary AllReduce Fallback Code by @caiomcbr in #391
• Fix mscclpp_benchmark by @Binyang2014 in #392
• Add cross threadblock barrier by @Binyang2014 in #383
• AllGather Executor Support in NCCL Interface by @caiomcbr in #393
• Providing reduce-scatter test support by @caiomcbr in #390
• Select algo according to json config by @Binyang2014 in #396
• Add connection events for NPKit by @yzygitzh in #386
• Revised ProxyChannel interfaces by @chhwang in #400
• Setup pipeline for mscclpp over nccl by @Binyang2014 in #401
• Exception Max Number Operation per Tb by @caiomcbr in #405
• Reduce memory usage for scratch buffer by @Binyang2014 in #403
• [Cherry-pick] Move pipeline to official org (#406) by @Binyang2014 in #416
• [Cherry-pick] trigger ci for release branches (#426) by @Binyang2014 in #427
• [Cherry-pick] Disable CuMemMap check for ROCm (#411) by @Binyang2014 in #424
• [Cherry-pick] NVLS support for NCCL API (#410) by @Binyang2014 in #425
• [Cherry-pick] Fix nccl-test failure issue (#421) by @Binyang2014 in #429

New Contributors

• @jeffra made their first contribution in #374

Full Changelog: v0.5.2...v0.6.0

MSCCL++ v0.5.2

What's Changed

• Add C++ executor test by @chhwang in #304
• Cumulative Updates by @Binyang2014 in #309
• Add NPKit GPU event support by @yzygitzh in #310
• Fix NPKit support for AMD by @yzygitzh in #312
• Add "packet type" option for executor test by @Binyang2014 in #313
• Add support for multicast reduce insruction by @roshandathathri in #316
• Update quickstart.md by @angelica-moreira in #314
• Simplify/improve barrier in AllReduce6 by @roshandathathri in #317
• Support NCCL APIs by @caiomcbr in #319
• Update allreduce_bench.py by @angelica-moreira in #318
• Separate NPKit CPU timestamp access from different blocks for AMD platform by @yzygitzh in #321
• AllReduce Kernel for Small Messages by @caiomcbr in #322
• Resolve clang++ warnings by @chhwang in #325
• Support to write packets via uint2 by @Binyang2014 in #327
• Double buffering for NCCL APIs by @caiomcbr in #324
• v0.5.2 by @chhwang in #328

New Contributors

• @angelica-moreira made their first contribution in #314
• @caiomcbr made their first contribution in #319

Full Changelog: v0.5.1...v0.5.2

MSCCL++ v0.5.1

What's Changed

• Upgrade gtest by @chhwang in #300
• Rename executor.cpp to executor_py.cpp by @chhwang in #301
• Fix assert declaration & add a compile test by @chhwang in #303
• Fix security issue by @Binyang2014 in #305
• v0.5.1 by @chhwang in #308

Full Changelog: v0.5.0...v0.5.1

MSCCL++ v0.5.0

What's Changed

• Fix a typo name by @chhwang in #286
• Add executor to execute schedule-plan file by @Binyang2014 in #283
• Allow binding allocated memory to NVLS multicast pointer by @roshandathathri in #290
• Seperate headers for GPU data types by @chhwang in #291
• Refactoring NVLS interfaces by @chhwang in #293
• Include GPU data types only for kernel code by @chhwang in #292
• Ethernet support by @chhwang in #284
• Resolve multi-nodes test failure issue by @Binyang2014 in #295
• Move pipeline to Azure org by @Binyang2014 in #296
• Optimized the execution kernel by @Binyang2014 in #294
• Allow obtaining cuda stream handle from PyTorch stream when launching kernel by @aashaka in #297
• v0.5.0 by @chhwang in #298

New Contributors

• @roshandathathri made their first contribution in #290

Full Changelog: v0.4.3...v0.5.0

MSCCL++ v0.4.3

What's Changed

• Add optional prefix to installation paths by @chhwang in #235
• Fix #235 by @chhwang in #239
• Check nvidia_peermem during runtime by @chhwang in #234
• Do not check value of __HIP_PLATFORM_AMD__ by @chhwang in #240
• Fix crash in static variable deconstructor by @Binyang2014 in #238
• Update interface to let user change fifo size by @Binyang2014 in #243
• Mask each fields of the trigger by @chhwang in #244
• Minor improvement on device syncer by @chhwang in #231
• remove make pylib-copy command by @Binyang2014 in #249
• Increase MSCCLPP_BITS_REGMEM_HANDLE to 9 by @aashaka in #251
• Add putWithSignal() latency tests by @chhwang in #246
• NVLS support. by @saeedmaleki in #250
• Fix wrong offset calculation by @chhwang in #257
• Fix NVLS support by @chhwang in #258
• Allow MSCCL++ CommGroup to take PyTorch tensors in args by @aashaka in #255
• Fix multi-nodes test failure by @Binyang2014 in #262
• Allow semaphores and memory to be registered separately in ProxyService by @aashaka in #264
• Remove cuda-python from project by @Binyang2014 in #245
• Fix the comm.py for nvls by @saeedmaleki in #267
• New packet format & optimizations by @chhwang in #256
• Fix multi-node ci pipeline by @Binyang2014 in #272
• add launch_bounds for mscclpp_test by @Binyang2014 in #273
• Fix bootstrapping mechanism by @chhwang in #278
• v0.4.3 by @chhwang in #279

New Contributors

• @aashaka made their first contribution in #251

Full Changelog: v0.4.2...v0.4.3

MSCCL++ v0.4.2

What's Changed

• Include cstdint in packet_device.hpp by @chhwang in #233
• Fix & improve perf for ROCm by @chhwang in #232
• v0.4.2 by @chhwang in #236

Full Changelog: v0.4.1...v0.4.2

MSCCL++ v0.4.1

What's Changed

• Fix performance downgrade issue & update doc by @Binyang2014 in #229
• Add a documentation issue template by @chhwang in #230

Full Changelog: v0.4.0...v0.4.1

MSCCL++ v0.4.0

• Add Python benchmark
• Update documentation
• Add ROCm support
• Bug fixes

See details from #160.

MSCCL++ v0.3.0

• Updated interfaces
• Add Python bindings and interfaces
• Add Python unit tests
• Add more configurable parameters
• Add a new single-node AllReduce kernel
• Fix bugs

See details from #89.

Full Changelog: v0.2.0...v0.3.0

MSCCL++ v0.2.0

Communication Features and Interfaces

GPU-side communication interfaces (DeviceChannel)

1. • Proxy-based Interfaces: ProxyChannel (#66)
2. • In-SM Copy Interfaces: SmChannel (#55)
3. • Packet Copy Interfaces: putPackets(), getPackets(), signalPacket() (#85, #90, #102)

Host-side interfaces

1. • Bootstrap: fix socket performance issue & bugs (#92, #100, #113)
2. • Communicator: implement (#66)

Transports support

1. • NVLink: implement (#66)
2. • InfiniBand: implement (#66)
3. • InfiniBand: tackle memory consistency issues (#96)

Performance Optimization

1. • Throughput: pass AllGather perf qualification (#77)
2. • Throughput: pass AllReduce perf qualification (#83, #90)
3. • Throughput: pass AllToAll perf qualification (#87)
4. • Latency: pass AllReduce perf qualification (#85, #90)
5. • Latency: pass 2-node AllReduce perf qualification (#109, #118)

Development Pipeline

1. • Unit Tests: cover all interfaces (#81, #91)
2. • mscclpp-test: add AllGather (#77)
3. • mscclpp-test: add AllReduce (#83)
4. • mscclpp-test: add AllToAll (#87)
5. • CI: lint, spelling, CodeQL (#79)
6. • CI: unit test (#81)
7. • CI: mscclpp-test (#93, #103)
8. • Package: publish Docker images (#104)

Documents

1. • Doxygen: add configuration (#72)
2. • README: enhance details (#88)
3. • License: add license comments on all files (#106)
4. • Code: cleanup & comments (#86, #119)

Full Changelog: https://github.com/microsoft/mscclpp/commits/v0.2.0